Awesome

GIBSON ENVIRONMENT for Embodied Active Agents with Real-World Perception

You shouldn't play video games all day, so shouldn't your AI! We built a virtual environment simulator, Gibson, that offers real-world experience for learning perception.

<img src=misc/ui.gif width="600">

Summary: Perception and being active (i.e. having a certain level of motion freedom) are closely tied. Learning active perception and sensorimotor control in the physical world is cumbersome as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given a fruitful rise to learning in the simulation which consequently casts a question on transferring to real-world. We developed Gibson environment with the following primary characteristics:

I. being from the real-world and reflecting its semantic complexity through virtualizing real spaces,
II. having a baked-in mechanism for transferring to real-world (Goggles function), and
III. embodiment of the agent and making it subject to constraints of space and physics via integrating a physics engine (Bulletphysics).

Naming: Gibson environment is named after James J. Gibson, the author of "Ecological Approach to Visual Perception", 1979. “We must perceive in order to move, but we must also move in order to perceive” – JJ Gibson

Please see the website (http://gibsonenv.stanford.edu/) for more technical details. This repository is intended for distribution of the environment and installation/running instructions.

Paper

"Gibson Env: Real-World Perception for Embodied Agents", in CVPR 2018 [Spotlight Oral].

Release

This is the 0.3.1 release. Bug reports, suggestions for improvement, as well as community developments are encouraged and appreciated. change log file.

Database

The full database includes 572 spaces and 1440 floors and can be downloaded here. A diverse set of visualizations of all spaces in Gibson can be seen here. To make the core assets download package lighter for the users, we include a small subset (39) of the spaces. Users can download the rest of the spaces and add them to the assets folder. We also integrated Stanford 2D3DS and Matterport 3D as separate datasets if one wishes to use Gibson's simulator with those datasets (access here).

Table of contents

• Installation
  • Quick Installation (docker)
  • Building from source
  • Uninstalling
• Quick Start
  • Gibson FPS
  • Web User Interface
  • Rendering Semantics
  • Robotic Agents
  • ROS Configuration
• Coding your RL agent
• Environment Configuration
• Goggles: transferring the agent to real-world
• Citation

Installation

Installation Method

There are two ways to install gibson, A. using our docker image (recommended) and B. building from source.

System requirements

The minimum system requirements are the following:

For docker installation (A):

• Ubuntu 16.04
• Nvidia GPU with VRAM > 6.0GB
• Nvidia driver >= 384
• CUDA >= 9.0, CuDNN >= v7

For building from the source(B):

• Ubuntu >= 14.04
• Nvidia GPU with VRAM > 6.0GB
• Nvidia driver >= 375
• CUDA >= 8.0, CuDNN >= v5

Download data

First, our environment core assets data are available here. You can follow the installation guide below to download and set up them properly. gibson/assets folder stores necessary data (agent models, environments, etc) to run gibson environment. Users can add more environments files into gibson/assets/dataset to run gibson on more environments. Visit the database readme for downloading more spaces. Please sign the license agreement before using Gibson's database.

A. Quick installation (docker)

We use docker to distribute our software, you need to install docker and nvidia-docker2.0 first.

Run docker run --runtime=nvidia --rm nvidia/cuda nvidia-smi to verify your installation.

You can either 1. pull from our docker image (recommended) or 2. build your own docker image.

1. Pull from our docker image (recommended)

# download the dataset from https://storage.googleapis.com/gibson_scenes/dataset.tar.gz
docker pull xf1280/gibson:0.3.1
xhost +local:root
docker run --runtime=nvidia -ti --rm -e DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix -v <host path to dataset folder>:/root/mount/gibson/gibson/assets/dataset xf1280/gibson:0.3.1

2. Build your own docker image

git clone https://github.com/StanfordVL/GibsonEnv.git
cd GibsonEnv
./download.sh # this script downloads assets data file and decompress it into gibson/assets folder
docker build . -t gibson ### finish building inside docker, note by default, dataset will not be included in the docker images
xhost +local:root ## enable display from docker

If the installation is successful, you should be able to run docker run --runtime=nvidia -ti --rm -e DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix -v <host path to dataset folder>:/root/mount/gibson/gibson/assets/dataset gibson to create a container. Note that we don't include dataset files in docker image to keep our image slim, so you will need to mount it to the container when you start a container.

Notes on deployment on a headless server

Gibson Env supports deployment on a headless server and remote access with x11vnc. You can build your own docker image with the docker file Dockerfile as above. Instructions to run gibson on a headless server (requires X server running):

1. Install nvidia-docker2 dependencies following the starter guide. Install x11vnc with sudo apt-get install x11vnc.
2. Have xserver running on your host machine, and run x11vnc on DISPLAY :0.
3. docker run --runtime=nvidia -ti --rm -e DISPLAY -v /tmp/.X11-unix/X0:/tmp/.X11-unix/X0 -v <host path to dataset folder>:/root/mount/gibson/gibson/assets/dataset <gibson image name>
4. Run gibson with python <gibson example or training> inside docker.
5. Visit your host:5900 and you should be able to see the GUI.

If you don't have X server running, you can still run gibson, see this guide for more details.

B. Building from source

If you don't want to use our docker image, you can also install gibson locally. This will require some dependencies to be installed.

First, make sure you have Nvidia driver and CUDA installed. If you install from source, CUDA 9 is not necessary, as that is for nvidia-docker 2.0. Then, let's install some dependencies:

apt-get update 
apt-get install libglew-dev libglm-dev libassimp-dev xorg-dev libglu1-mesa-dev libboost-dev \
		mesa-common-dev freeglut3-dev libopenmpi-dev cmake golang libjpeg-turbo8-dev wmctrl \
		xdotool libzmq3-dev zlib1g-dev

Install required deep learning libraries: Using python3.5 is recommended. You can create a python3.5 environment first.

conda create -n py35 python=3.5 anaconda 
source activate py35 # the rest of the steps needs to be performed in the conda environment
conda install -c conda-forge opencv
pip install http://download.pytorch.org/whl/cu90/torch-0.3.1-cp35-cp35m-linux_x86_64.whl 
pip install torchvision==0.2.0
pip install tensorflow==1.3

Clone the repository, download data and build

git clone https://github.com/StanfordVL/GibsonEnv.git
cd GibsonEnv
./download.sh # this script downloads assets data file and decompress it into gibson/assets folder
./build.sh build_local ### build C++ and CUDA files
pip install -e . ### Install python libraries

Install OpenAI baselines if you need to run the training demo.

git clone https://github.com/fxia22/baselines.git
pip install -e baselines

Uninstalling

Uninstall gibson is easy. If you installed with docker, just run docker images -a | grep "gibson" | awk '{print $3}' | xargs docker rmi to clean up the image. If you installed from source, uninstall with pip uninstall gibson

Quick Start

First run xhost +local:root on your host machine to enable display. You may need to run export DISPLAY=:0 first. After getting into the docker container with docker run --runtime=nvidia -ti --rm -e DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix -v <host path to dataset folder>:/root/mount/gibson/gibson/assets/dataset gibson, you will get an interactive shell. Now you can run a few demos.

If you installed from source, you can run those directly using the following commands without using docker.

python examples/demo/play_husky_nonviz.py ### Use ASWD keys on your keyboard to control a car to navigate around Gates building

<img src=misc/husky_nonviz.png width="600">

You will be able to use ASWD keys on your keyboard to control a car to navigate around Gates building. A camera output will not be shown in this particular demo.

python examples/demo/play_husky_camera.py ### Use ASWD keys on your keyboard to control a car to navigate around Gates building, while RGB and depth camera outputs are also shown.

<img src=misc/husky_camera.png width="600">

You will able to use ASWD keys on your keyboard to control a car to navigate around Gates building. You will also be able to see the RGB and depth camera outputs.

python examples/train/train_husky_navigate_ppo2.py ### Use PPO2 to train a car to navigate down the hallway in Gates building, using visual input from the camera.

<img src=misc/husky_train.png width="800"> By running this command you will start training a husky robot to navigate in Gates building and go down the corridor with RGBD input. You will see some RL related statistics in the terminal after each episode.

python examples/train/train_ant_navigate_ppo1.py ### Use PPO1 to train an ant to navigate down the hallway in Gates building, using visual input from the camera.

<img src=misc/ant_train.png width="800"> By running this command you will start training an ant to navigate in Gates building and go down the corridor with RGBD input. You will see some RL related statistics in the terminal after each episode.

Gibson Framerate

Below is Gibson Environment's framerate benchmarked on different platforms. Please refer to fps branch for the code to reproduce the results.

We also tested on <code>Intel I7 7700 + NVIDIA GeForce GTX 1070Ti</code> and <code>Tested on Intel I7 6580k + NVIDIA GTX 1080Ti</code> platforms. The FPS difference is within 10% on each task.

<img src=misc/mpi_fps.png width="600">

Web User Interface

When running Gibson, you can start a web user interface with python gibson/utils/web_ui.py python gibson/utils/web_ui.py 5552. This is helpful when you cannot physically access the machine running gibson or you are running on a headless cloud environment. You need to change mode in configuration file to web_ui to use the web user interface.

<img src=misc/web_ui.png width="600">

Rendering Semantics

<img src=misc/instance_colorcoding_semantics.png width="600">

Gibson can provide pixel-wise frame-by-frame semantic masks when the model is semantically annotated. As of now we have incorporated models from Stanford 2D-3D-Semantics Dataset and Matterport 3D for this purpose. You can access them within Gibson here. We refer you to the original dataset's reference for the list of their semantic classes and annotations.

For detailed instructions of rendering semantics in Gibson, see semantic instructions. As one example in the starter dataset that comes with installation, space7 includes Stanford 2D-3D-Semantics style annotation.

Robotic Agents

Gibson provides a base set of agents. See videos of these agents and their corresponding perceptual observation here. <img src=misc/agents.gif>

To enable (optionally) abstracting away low-level control and robot dynamics for high-level tasks, we also provide a set of practical and ideal controllers for each agent.

Starter Code

More demonstration examples can be found in examples/demo folder

More training code can be found in examples/train folder.

ROS Configuration

We provide examples of configuring Gibson with ROS here. We use turtlebot as an example, after a policy is trained in Gibson, it requires minimal changes to deploy onto a turtlebot. See README for more details.

Coding Your RL Agent

You can code your RL agent following our convention. The interface with our environment is very simple (see some examples in the end of this section).

First, you can create an environment by creating an instance of classes in gibson/core/envs folder.

env = AntNavigateEnv(is_discrete=False, config = config_file)

Then do one step of the simulation with env.step. And reset with env.reset()

obs, rew, env_done, info = env.step(action)

obs gives the observation of the robot. It is a dictionary with each component as a key value pair. Its keys are specified by user inside config file. E.g. obs['nonviz_sensor'] is proprioceptive sensor data, obs['rgb_filled'] is rgb camera data.

rew is the defined reward. env_done marks the end of one episode, for example, when the robot dies. info gives some additional information of this step; sometimes we use this to pass additional non-visual sensor values.

We mostly followed OpenAI gym convention when designing the interface of RL algorithms and the environment. In order to help users start with the environment quicker, we provide some examples at examples/train. The RL algorithms that we use are from openAI baselines with some adaptation to work with hybrid visual and non-visual sensory data. In particular, we used PPO and a speed optimized version of PPO.

Environment Configuration

Each environment is configured with a yaml file. Examples of yaml files can be found in examples/configs folder. Parameters for the file is explained below. For more informat specific to Bullet Physics engine, you can see the documentation here.

Making Your Customized Environment

Gibson provides a set of methods for you to define your own environments. You can follow the existing environments inside gibson/core/envs.

Goggles: transferring the agent to real-world

Gibson includes a baked-in domain adaptation mechanism, named Goggles, for when an agent trained in Gibson is going to be deployed in real-world (i.e. operate based on images coming from an onboard camera). The mechanisms is essentially a learned inverse function that alters the frames coming from a real camera to what they would look like if they were rendered via Gibson, and hence, disolve the domain gap.

<img src=http://gibson.vision/public/img/figure4.jpg width="600">

More details: With all the imperfections in point cloud rendering, it has been proven difficult to get completely photo-realistic rendering with neural network fixes. The remaining issues make a domain gap between the synthesized and real images. Therefore, we formulate the rendering problem as forming a joint space ensuring a correspondence between rendered and real images, rather than trying to (unsuccessfully) render images that are identical to real ones. This provides a deterministic pathway for traversing across these domains and hence undoing the gap. We add another network "u" for target image (I_t) and define the rendering loss to minimize the distance between f(I_s) and u(I_t), where "f" and "I_s" represent the filler neural network and point cloud rendering output, respectively (see the loss in above figure). We use the same network structure for f and u. The function u(I) is trained to alter the observation in real-world, I_t, to look like the corresponding I_s and consequently dissolve the gap. We named the u network goggles, as it resembles corrective lenses for the agent for deployment in real-world. Detailed formulation and discussion of the mechanism can be found in the paper. You can download the function u and apply it when you deploy your trained agent in real-world.

In order to use goggle, you will need preferably a camera with depth sensor, we provide an example here for Kinect. The trained goggle functions are stored in assets/unfiller_{resolution}.pth, and each one is paired with one filler function. You need to use the correct one depending on which filler function is used. If you don't have a camera with depth sensor, we also provide an example for RGB only here.

Citation

If you use Gibson Environment's software or database, please cite:

@inproceedings{xiazamirhe2018gibsonenv,
  title={Gibson {Env}: real-world perception for embodied agents},
  author={Xia, Fei and R. Zamir, Amir and He, Zhi-Yang and Sax, Alexander and Malik, Jitendra and Savarese, Silvio},
  booktitle={Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on},
  year={2018},
  organization={IEEE}
}